Keigo and AI

Also called: Japanese honorifics AI, keigo generation

Keigo is the Japanese honorific system that encodes the relationship between speaker, listener and referent, and AI assistants handle it unreliably because the correct level is determined by social context the model is not given rather than by the sentence being translated.

The error is rarely grammatical; it is a wrong relationship, which reads as rude rather than as a mistake.

A model can inflect sonkeigo, kenjougo and teineigo correctly and still be wrong, because choosing among them requires knowing who outranks whom, how close the parties are, and whether the writer is speaking for themselves or for their company. None of that is in the source sentence.

The consequence is asymmetric risk. English written slightly too formally is mildly odd. Japanese written at the wrong level of keigo damages the relationship the message exists to maintain, and the writer often cannot see the damage in their own second language.

DeterminantExampleAvailable to the model by default?
Relative seniorityWriting to a client's director versus a peerNo
In-group or out-groupReferring to your own manager to an outsiderNo
Prior correspondence registerThe level the counterpart used firstOnly if the thread is supplied
Industry conventionFinance versus startup normsPartially, and unevenly
Grammatical forms themselvesirassharu, itadaku, desu/masuYes
What determines the correct level, and whether a model is given it

Frequently asked questions

Why does AI get keigo wrong if it knows the grammar?
Because the grammar is the easy half. The hard half is selecting the level, which depends on social facts about the sender and recipient that are not present in the text being written.
Does a writing profile help with keigo?
It helps where the author has an established habit — the level they actually use with a given class of counterpart — because the profile carries that habit instead of asking the model to infer it. It does not remove the need to state who the recipient is.
Is this specific to Japanese?
The severity is, but the pattern is not. French tu and vous and German du and Sie pose the same problem in a smaller space, and the same fix applies: supply the relationship, do not hope the model guesses it.

Read further

Sources and how to cite this page

Cite this page

MyWritingTwin. “Keigo and AI.” Last updated 2026-09-15. https://www.mywritingtwin.com/glossary/keigo-and-ai